Skip to contentSkip to conversation

bpf, arm64: Call bpf_jit_binary_pack_finalize() in bpf_jit_free()

PATCH In review Hengqi Chen 4 participants 6 messages Aug 28, 2025 – Sep 17, 2025

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf, arm64: Call bpf_jit_binary_pack_finalize() in bpf_jit_free()

The current implementation seems incorrect and does NOT match the comment above, use bpf_jit_binary_pack_finalize() instead.

Fixes 1dad391daef1 ("bpf, arm64: use bpf_prog_pack for memory management")
---
 arch/arm64/net/bpf_jit_comp.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
arch/arm64/net/bpf_jit_comp.c +1 -2
index 52ffe115a8c4..4ef9b7b8fb40 100644
@@ -3064,8 +3064,7 @@ void bpf_jit_free(struct bpf_prog *prog)
 		 * before freeing it.
 		 */
 		if (jit_data) {
-			bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
-					   sizeof(jit_data->header->size));
+			bpf_jit_binary_pack_finalize(jit_data->ro_header, jit_data->header);
 			kfree(jit_data);
 		}
 		prog->bpf_func -= cfi_get_offset();
Signature · 1 line
2.27.0
maintainer
#1 lore raw
Quoted 14 lines · Hengqi Chen

The current implementation seems incorrect and does NOT match the comment above, use bpf_jit_binary_pack_finalize() instead.

Fixes 1dad391daef1 ("bpf, arm64: use bpf_prog_pack for memory management")
---
 arch/arm64/net/bpf_jit_comp.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
arch/arm64/net/bpf_jit_comp.c +0 -0
index 52ffe115a8c4..4ef9b7b8fb40 100644

arch/arm64/net/bpf_jit_comp.c void bpf_jit_free(struct bpf_prog *prog) · line 3068

Quoted 7 lines
arch/arm64/net/bpf_jit_comp.c +1 -2
@@ -3064,8 +3064,7 @@ void bpf_jit_free(struct bpf_prog *prog)
 		 * before freeing it.
 		 */
 		if (jit_data) {
-			bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
-					   sizeof(jit_data->header->size));
+			bpf_jit_binary_pack_finalize(jit_data->ro_header, jit_data->header);
 			kfree(jit_data);

Thanks for this patch!

So, this is fixing a bug because bpf_jit_binary_pack_finalize() will do kvfree(rw_header); but without it currently, jit_data->header is never freed.

But I think we shouldn't use bpf_jit_binary_pack_finalize() here as it copies the whole rw_header to ro_header using bpf_arch_text_copy() which is an expensive operation (patch_map/unmap in loop + flush_icache_range()) and not needed here because we are going to free ro_header anyway.

We only need to copy jit_data->header->size to jit_data->ro_header->size because this size is later used by bpf_jit_binary_pack_free(), see comment above bpf_jit_binary_pack_free().

How I suggest we should fix the code and the comment:

-- >8 --

arch/arm64/net/bpf_jit_comp.c +4 -2
index 5083886d6e66b..cb4c50eeada13 100644
@@ -3093,12 +3093,14 @@ void bpf_jit_free(struct bpf_prog *prog)
 
                /*
                 * If we fail the final pass of JIT (from jit_subprogs),
-                * the program may not be finalized yet. Call finalize here
-                * before freeing it.
+                * the program may not be finalized yet. Copy the header size
+                * from rw_header to ro_header before freeing the ro_header
+                * with bpf_jit_binary_pack_free().
                 */
                if (jit_data) {
                        bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
                                           sizeof(jit_data->header->size));
+                       kvfree(jit_data->header);
                        kfree(jit_data);
                }
                prog->bpf_func -= cfi_get_offset();

-- 8< --

Song,

Do you think this optimization is worth it or should we just call bpf_jit_binary_pack_finalize() here like this patch is doing?

Thanks,
Puranjay

reviewer
#2 lore raw

Sorry for the late reply.

On Thu, Aug 28, 2025 at 5:10 AM Puranjay Mohan <[email protected]> wrote:
[...]

Quoted 44 lines

Thanks for this patch!

So, this is fixing a bug because bpf_jit_binary_pack_finalize() will do kvfree(rw_header); but without it currently, jit_data->header is never freed.

But I think we shouldn't use bpf_jit_binary_pack_finalize() here as it copies the whole rw_header to ro_header using bpf_arch_text_copy() which is an expensive operation (patch_map/unmap in loop + flush_icache_range()) and not needed here because we are going to free ro_header anyway.

We only need to copy jit_data->header->size to jit_data->ro_header->size because this size is later used by bpf_jit_binary_pack_free(), see comment above bpf_jit_binary_pack_free().

How I suggest we should fix the code and the comment:

-- >8 --

arch/arm64/net/bpf_jit_comp.c +4 -3
index 5083886d6e66b..cb4c50eeada13 100644
@@ -3093,12 +3093,14 @@ void bpf_jit_free(struct bpf_prog *prog)
 
                /*
                 * If we fail the final pass of JIT (from jit_subprogs),
-                * the program may not be finalized yet. Call finalize here
-                * before freeing it.
+                * the program may not be finalized yet. Copy the header size
+                * from rw_header to ro_header before freeing the ro_header
+                * with bpf_jit_binary_pack_free().
                 */
                if (jit_data) {
                        bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
                                           sizeof(jit_data->header->size));
+                       kvfree(jit_data->header);
                        kfree(jit_data);
                }
                prog->bpf_func -= cfi_get_offset();
 
-- 8< --
Quoted 4 lines

Song,

Do you think this optimization is worth it or should we just call bpf_jit_binary_pack_finalize() here like this patch is doing?

This is a good optimization. However, given this is not a hot path, I don't have a strong preference either way. At the moment, most other architectures use bpf_jit_binary_pack_finalize(), so it is good to just use bpf_jit_binary_pack_finalize and keep the logic consistent.

In the longer term, we can consider refactoring bpf_jit_free so that multiple architectures can share code. After this patch, bpf_jit_free for x86_64 and arm64 are very similar to each other. There are likely some opportunities to reduce code duplications.

Thanks,
Song

maintainer Acked-by
#3 lore raw
Quoted 50 lines · Song Liu

Sorry for the late reply.

On Thu, Aug 28, 2025 at 5:10 AM Puranjay Mohan <[email protected]> wrote:
[...]

Quoted 46 lines

Thanks for this patch!

So, this is fixing a bug because bpf_jit_binary_pack_finalize() will do kvfree(rw_header); but without it currently, jit_data->header is never freed.

But I think we shouldn't use bpf_jit_binary_pack_finalize() here as it copies the whole rw_header to ro_header using bpf_arch_text_copy() which is an expensive operation (patch_map/unmap in loop + flush_icache_range()) and not needed here because we are going to free ro_header anyway.

We only need to copy jit_data->header->size to jit_data->ro_header->size because this size is later used by bpf_jit_binary_pack_free(), see comment above bpf_jit_binary_pack_free().

How I suggest we should fix the code and the comment:

-- >8 --

arch/arm64/net/bpf_jit_comp.c +4 -3
index 5083886d6e66b..cb4c50eeada13 100644
@@ -3093,12 +3093,14 @@ void bpf_jit_free(struct bpf_prog *prog)
 
                /*
                 * If we fail the final pass of JIT (from jit_subprogs),
-                * the program may not be finalized yet. Call finalize here
-                * before freeing it.
+                * the program may not be finalized yet. Copy the header size
+                * from rw_header to ro_header before freeing the ro_header
+                * with bpf_jit_binary_pack_free().
                 */
                if (jit_data) {
                        bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
                                           sizeof(jit_data->header->size));
+                       kvfree(jit_data->header);
                        kfree(jit_data);
                }
                prog->bpf_func -= cfi_get_offset();
 
-- 8< --

Song,

Quoted 8 lines
Quoted 2 lines

Do you think this optimization is worth it or should we just call bpf_jit_binary_pack_finalize() here like this patch is doing?

This is a good optimization. However, given this is not a hot path, I don't have a strong preference either way. At the moment, most other architectures use bpf_jit_binary_pack_finalize(), so it is good to just use bpf_jit_binary_pack_finalize and keep the logic consistent.

So, in that case we can merge this patch.

Thanks,
Puranjay

#4 lore raw
Quoted 62 lines · Puranjay Mohan
Quoted 50 lines · Song Liu

Sorry for the late reply.

On Thu, Aug 28, 2025 at 5:10 AM Puranjay Mohan <[email protected]> wrote:
[...]

Quoted 46 lines

Thanks for this patch!

So, this is fixing a bug because bpf_jit_binary_pack_finalize() will do kvfree(rw_header); but without it currently, jit_data->header is never freed.

But I think we shouldn't use bpf_jit_binary_pack_finalize() here as it copies the whole rw_header to ro_header using bpf_arch_text_copy() which is an expensive operation (patch_map/unmap in loop + flush_icache_range()) and not needed here because we are going to free ro_header anyway.

We only need to copy jit_data->header->size to jit_data->ro_header->size because this size is later used by bpf_jit_binary_pack_free(), see comment above bpf_jit_binary_pack_free().

How I suggest we should fix the code and the comment:

-- >8 --

arch/arm64/net/bpf_jit_comp.c +4 -3
index 5083886d6e66b..cb4c50eeada13 100644
@@ -3093,12 +3093,14 @@ void bpf_jit_free(struct bpf_prog *prog)
 
                /*
                 * If we fail the final pass of JIT (from jit_subprogs),
-                * the program may not be finalized yet. Call finalize here
-                * before freeing it.
+                * the program may not be finalized yet. Copy the header size
+                * from rw_header to ro_header before freeing the ro_header
+                * with bpf_jit_binary_pack_free().
                 */
                if (jit_data) {
                        bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
                                           sizeof(jit_data->header->size));
+                       kvfree(jit_data->header);
                        kfree(jit_data);
                }
                prog->bpf_func -= cfi_get_offset();
 
-- 8< --

Song,

Quoted 8 lines
Quoted 2 lines

Do you think this optimization is worth it or should we just call bpf_jit_binary_pack_finalize() here like this patch is doing?

This is a good optimization. However, given this is not a hot path, I don't have a strong preference either way. At the moment, most other architectures use bpf_jit_binary_pack_finalize(), so it is good to just use bpf_jit_binary_pack_finalize and keep the logic consistent.

Quoted 3 lines

So, in that case we can merge this patch.

It's out of patchwork.
Hengqi,
pls repost.

#5 lore raw
Quoted 64 lines · Alexei Starovoitov
Quoted 62 lines · Puranjay Mohan
Quoted 58 lines · Song Liu

Sorry for the late reply.

On Thu, Aug 28, 2025 at 5:10 AM Puranjay Mohan <[email protected]> wrote:
[...]

Quoted 44 lines

Thanks for this patch!

So, this is fixing a bug because bpf_jit_binary_pack_finalize() will do kvfree(rw_header); but without it currently, jit_data->header is never freed.

But I think we shouldn't use bpf_jit_binary_pack_finalize() here as it copies the whole rw_header to ro_header using bpf_arch_text_copy() which is an expensive operation (patch_map/unmap in loop + flush_icache_range()) and not needed here because we are going to free ro_header anyway.

We only need to copy jit_data->header->size to jit_data->ro_header->size because this size is later used by bpf_jit_binary_pack_free(), see comment above bpf_jit_binary_pack_free().

How I suggest we should fix the code and the comment:

-- >8 --

arch/arm64/net/bpf_jit_comp.c +4 -3
index 5083886d6e66b..cb4c50eeada13 100644
@@ -3093,12 +3093,14 @@ void bpf_jit_free(struct bpf_prog *prog)
 
                /*
                 * If we fail the final pass of JIT (from jit_subprogs),
-                * the program may not be finalized yet. Call finalize here
-                * before freeing it.
+                * the program may not be finalized yet. Copy the header size
+                * from rw_header to ro_header before freeing the ro_header
+                * with bpf_jit_binary_pack_free().
                 */
                if (jit_data) {
                        bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
                                           sizeof(jit_data->header->size));
+                       kvfree(jit_data->header);
                        kfree(jit_data);
                }
                prog->bpf_func -= cfi_get_offset();
 
-- 8< --
Quoted 4 lines

Song,

Do you think this optimization is worth it or should we just call bpf_jit_binary_pack_finalize() here like this patch is doing?

This is a good optimization. However, given this is not a hot path, I don't have a strong preference either way. At the moment, most other architectures use bpf_jit_binary_pack_finalize(), so it is good to just use bpf_jit_binary_pack_finalize and keep the logic consistent.

Quoted 7 lines
Quoted 3 lines

So, in that case we can merge this patch.

It's out of patchwork.
Hengqi,
pls repost.

OK, just resend.